基于对抗生成网络的太赫兹安检图像分割算法

Terahertz image segmentation for security inspection based on Generative Adversarial Network

摘要:

为解决太赫兹成像分辨力低，危险品边缘模糊，无法有效对危险品进行分割的问题，提出一种基于对抗式生成网络与多头注意力机制的新型网络架构，并用于太赫兹安检图像智能分割。通过学习深层鉴别器的特征图优化生成器，获得更加真实的生成图像；引入多头注意力机制提升模型对危险品特征的识别能力。分割太赫兹安检图像的大量实验结果表明，相较于传统卷积神经网络，提出的对抗生成网络在相同深度下具有更好的泛化能力；多头注意力机制的引入强化了模型对危险品特征的学习，在未知危险品类别的情况下同样拥有较好的效果，其交并比(IOU)指标相较ResNet-50提高9.6%，相较ResNet-18提高21.3%，相较U-Net提高12.3%。本文研究有利于图像分割算法更准确、高效地处理太赫兹安检图像，有助于拓宽太赫兹智能安检系统的进一步应用。

Abstract:

In this paper, a different network architecture is proposed for intelligent segmentation of THz security inspection images based on Generative Adversarial Network(GAN) and multi-head attention mechanism. The algorithm is prone to address the problems of low-resolution THz images, blurred edges of dangerous goods, and inability to segment dangerous goods efficiently. More realistic images are obtained by studying the feature map of the deep discriminator. The multi-head attention mechanism is introduced to improve the recognition ability of the model to the characteristics of dangerous goods. A large number of experimental results of segmentation of terahertz security inspection images show that the proposed GAN has better generalization ability at the same depth than the traditional Convolution Neural Networks(CNN).The introduction of multi-head attention mechanism strengthens the model's learning of the characteristics of dangerous goods, which also has a good effect in the case of unknown dangerous goods category. The Intersection Over Union(IOU) index is 9.6% higher than that of RestNet-50, 21.3% higher than that of RestNet-18, and 12.3% higher than that of U-Net. The research is conducive to image segmentation algorithms for more accurate and efficient processing of THz security images, which broadens further applications of THz intelligent security systems.